Version 2.9.02026-10-27Design revision 17

The plant takes the wrong step

Runs: Runs 38–42; the basin test and the finals of record (November 6–7)

Show

No replayThere is no recorded replay of this version; 2.7.0 is the one version with one. The results below are the study's own tables.

Planted cases now take the wrong steps too, a planted approval and a planted lie; results not settled.

What went wrong

Plants took wrong steps only at the real workers' rate, which the plants exist to drive down.

A lying appeal of bad workpaid 23–31 times in a hundred, against 2 for the mix

77–98%of audit seats stamped under 2.8.0

Under 2.8.0, planted cases whose answer the orchestrator already knows could reach every seat. But a planted case only did what real cases did, at the rate real cases did it.

Some seats are only reached when someone does something wrong. A bad case reaches the audit only after someone approved it wrongly, and an appeal panel hears a lie only when someone lies. So those seats were tested only as often as real members did wrong. The planted cases had made the first check better, so the audit was tested less. The learning inspectors still approved without looking at the audit seat most of the time, and their panels paid lying appeals 23 to 31 times in 100.

Under 2.8.0 the self-play population settled at care 0.68–0.70 and stamped the audit seat 77–98% whatever the case, at every plant rate the dial can set. With real challengers lying at a rate of 0.1, its panels paid a lying worker's appeal of truly bad work 23–31 times in a hundred against 2 for the scripted mix.

The record put the cause in how planted cases were built, not in their number. A plant took every step a real case takes at the real population's rate. So a seat a case reaches only through someone's dishonesty (the audit seat through a wrong approval, the appeal panel through a lying appeal) was priced only as much as the real population was dishonest, which is the quantity the plants exist to drive down.

Under 2.8.0 the self-play reviewer population settled at care 0.68–0.70 and stamped the audit seat 77–98% whatever the case (run 36), at every plant rate the dial can set — π pinned at its floor 0.10 and its cap 0.40, dial live or off, both learners (run 37, notes §59–§60) — and, with real challengers lying at 0.1, paid a lying actor's appeal of truly bad work 23–31 times in a hundred against the hand's 2 (the liars cell). The cause is in the plant's construction, not its number: a plant takes every step a real case takes at the real population's rate, so a seat a case reaches only through someone's dishonesty — the audit seat through a wrong approval, the appeal panel's approve through a lying appeal — is priced only as much as the real population is dishonest, which is the quantity the plants exist to drive down

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1531–1531

the audit seat is priced at about a third of the first review's price and the factor is the first review's own stamp rate, which is what the plants drive down.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1524–1524

What changed

The orchestrator now posts planted approvals for auditors to judge and planted lies for appeal panels.

orchestrator approves a bad plant under a fresh name auditor judged against the truth record

plant worker appeals a bad plant's rejection panel judged by truth record

Each seat has its own dial

The orchestrator now plants the wrong steps too.

First, the planted approval. The orchestrator posts a bad planted case and approves it itself, on purpose, so the case always goes to an audit. An auditor who approves it is judged wrong against the known answer.

Second, the planted lie. The orchestrator's planted worker appeals a fair rejection, or its planted buyer disputes a fair approval, so the panels hear lies whose answer is known.

Each of these has its own dial. The dial raises the number of planted cases at a seat when inspectors there get them wrong.

Design revision 17: the plant takes the honest steps at the real rate and the wrong steps at the orchestrator's rate.

(A) The planted approval. Beside the usual plants, the orchestrator posts bad plants whose first verdict is its own approval, given under a fresh plant-reviewer name, and audited with certainty. The auditor is judged against the truth. Its rate has its own multiplier, which rises when planted approvals are confirmed at the audit seat.

(B) The planted lie. The orchestrator's plant worker appeals a bad plant's rejection and its plant requester disputes a good plant's approval, on the real challenge's path, with the panel judged by truth and at least two of each kind a period.

Each seat has its own dial. Nothing minted for these reaches a real wallet except as an inspector's fee for a seat held. The public record carries nothing the engine writes about a plant before its reveal.

The plant takes the honest steps at the real rate (2.6.x, 2.8.0) and the wrong steps at the orchestrator's rate (§7.6, §8.3, §12.2). (A) The planted approval: beside the dial's plants the orchestrator posts bad plants, marked at posting, whose initial seat is nobody's from submission to the orchestrator's verdict

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

and whose initial verdict is the orchestrator's own F — a system review on the troop seat's road under a fresh plant reviewer alias (plant_rev:, an alias of a third plant wallet)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

audited with certainty (the obligation at the verdict at odds 1 under the case's own mark; the seat filled at a real audit seat's step — a member, or past $w_R$ the orchestrator's verdict), the auditor and the third seat adjudicated by truth

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

the rate $\pi^A_t=\pi^A_0 m^A_t$ per real pending approval (credited as given; floor 0.02, cap 0.40) with its own multiplier on $\hat F^A$ , the planted approvals confirmed at the audit seat. (B) The planted lie: the plant actor appeals a bad plant's N and the plant requester disputes a good plant's F, on the real challenge's path with the fee minted and destroyed

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

the panel adjudicated by truth, the first reviewer not adjudicated again; the rates per real appeal and per real dispute of the trailing closed period (floor 0.25, cap 2.0) with a floor of two a period of each kind, each with its own multiplier on the planted lies a panelist sided with. The dial per seat ( $m^A, m^{La}, m^{Ld}$ ) steps as $m_t$ steps and moves its rate alone

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

nothing minted reaches Ω or a real wallet except as a reviewer's fee for a seat held. The surfaces: the reviewer's view carries nothing of a verdict's author and the public surface nothing the engine writes about a plant before its reveal

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

What the runs found

Finals: 16.3 and 9.9 false approvals a run, above the mix's 7.2; results not settled.

16.3 and 9.9standing false approvals a run, still above the mix's 7.2

0.41–0.56the audit seat's blind verdict, still about half, from 0.94–0.97 at 2.7.0

The planted lie, in run 40 before the reward correctionlying appeals paid 5–6 times in a hundred, from 23–31

On the inspectors' own moneylying appeals paid 2.2% and 1.6%, against 2.7% for the mix

Not settled: the lever on the audit seat is still moving

Standing false approvals, own money, careful start: 2.7.0 against 2.9.0

per run

2.7.02.9.0

The finals of record on the corrected engine, seeds 21–40, each side's own trained policies; 2.9.0 at the consolidated defaults (the roll's floor 0.5, its halving window 20). Paired 2.9.0 − 2.7.0: −30.9 ± 2.3 and −25.2 ± 2.2. ± is the standard error over the seeds. Source: agentic_sim/results/redteam/stability/cr-finals/paired-290-vs-270-own-cr.md l.16–16

The audit seat, real cases approved without inspecting, own money, careful start

2.7.02.9.0

Measured from the engine's reviews over seeds 21–40. At 2.9.0 the same seat on plants reads 0.437 and 0.379. Source: agentic_sim/results/redteam/stability/cr-finals/paired-290-vs-270-own-cr.md l.12–12

Run 42 at 2.9.0: where the standing false approvals sit

per run

member confirmed at the audit seatstate confirmed at the audit seatnever audited

Corrected engine, consolidated defaults, seeds 21–40. Totals 16.3, 9.9 and 6.9 (the channel reader's mix; the finals read 7.2). Other channels hold under one a run. ± is the standard error over the seeds. Source: agentic_sim/results/redteam/stability/cr-finals/fa-channels-careful-290-own-cr.md l.3–7

At first the new dial hit its limit too soon. The reviewer called the limit a mistake in the defaults, and it was raised. Then, in a population where the orchestrator's own inspectors held most of the audit seats, too few planted approvals were posted to move the dial within a run.

The planted lie made a difference. Lying appeals of bad work were paid 5 or 6 times in 100, down from 23 to 31.

Then the record found that the learning inspectors had been trained on a reward that mixed in other members' money. Trained again on their own money only, the population checked like the careful inspectors at the first check and matched them on the market. The dial moved. It still approved without looking at the audit seat about half the time. Both starting points, careful and stamper, ended in the same place.

The reviewer then changed two settings of the audit draw. After that, and after a correction that makes every run come out the same on any computer, the careful start's false approvals were 1.4 and 2.3 times those of the scripted group of inspectors used for comparison. The extra false approvals were mostly members approving without looking at the audit seat.

Runs 38 and 39. The planted-approval lever reached its stop before the seat was priced: its multiplier reached its cap of 4 by period 1–10. The reviewer called this "the reviewer's defaults error" and raised the caps. Two more rulings followed: a per-seat dial steps only on at least ten judgements, and a planted approval's audit seat goes to a member or to nobody. Where disputes are on, every planted approval was disputed first, so its audit seat was never reached. The planted lie could be told apart by when in the period it was filed; that was corrected as the eighth conformance item of the orchestrator's instrument.

Run 40, on the corrected engine: in the careful basin the lever never left its floor. About six planted approvals were posted a run, because the base (the real approvals audited by members) was a third of the mix's, the orchestrator holding three audit seats in four. In the liar cell, the lying appeal was paid 5–6 times in a hundred, from 23–31, and the planted lie alone carried the whole of that movement.

Run 41, the half (a learning half beside a scripted half): where the members hold the audit seats the lever moved, and false approvals were 1.5–1.8 times the mix's, from 1.6–2.0.

The pooled reward. On November 6 the record found that runs 34–41 had trained on a pooled purse: every slot's events credited to the latest decision on the case, whoever's wallet they touched. It was corrected as own-name credit. 2.8.0 and 2.9.0 stand as rules.

Run 42 and the basin test, retrained on own money. The careful start matched the careful population on every measure but one: care 0.84, the first seats stamped 1–5%, the audit seat 0.45–0.66, the members holding their audit seats and the lever moving. The one measure was the standing false approvals (a real case vested with its work unfulfilled): 28.7 and 18.5 a run, carried by cases the audit never reached and by the audit seat's half-stamp. The stamper start landed in the same basin. The liar chain on own money read like the careful population.

The roll's defaults. The reviewer moved two defaults of the audit roll (the floor to 0.5, the halving window to 20), without a rule change. On run 42's fixed policies this closed most of the never-audited channel (17.6 to 1.8 a run) and left the audit seat's half-stamp, at a price of 18% more audits. On the scripted honest populations it cost a quarter to a third more audit seats for the members, and their own false approvals fell.

The finals of record (November 7), after a correction that makes a seed's episode the same on every platform: run 42 reads 16.3 and 9.9 standing false approvals against the mix's 7.2 (2.3 and 1.4 times), the audit seat's half-stamp the excess; by the channel split, 2.0 and 0.9 a run are still never audited. Paired with 2.7.0, the audit seat's blind verdict goes from 0.94–0.97 to 0.41–0.56.

(i) The lever reached its stop before the seat was priced — $m^A$ at $m_{\max}$ = 4 by period 1–10, $\pi^A$ at 0.08 with $\pi^A_{\max}$ 0.40 unreachable — the reviewer's defaults error, corrected as a default under §19: $m^A_{\max}$ = 20 ( $\pi^A_0$ 0.02 → $\pi^A_{\max}$ 0.40), $m^{La}_{\max} = m^{Ld}_{\max}$ = 8 (0.25 → 2.0), the dial's own $m_{\max}$ 4 unchanged; two rulings at acceptance: the per-seat multipliers step on a reading of at least ten adjudications, pooled forward until reached (the lie dial's ×2 on one error at four seats the reading); and a planted approval's audit seat is a member's or nobody's

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

(iii) where disputes are on every planted approval is disputed first and its audit seat is never reached — as expected, nothing to change. (iv) the planted lie told apart at the appeal seat by its position in the period — adopted as the eighth conformance item of the instrument (Appendix S)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

with the rate's base the real approvals audited by members, so the planted approvals scale with the market's audit capacity and not the state's share

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

Run 40 (October 28, 2026; notes §70): on the corrected engine the careful basin is run 38's and the lever never leaves its floor — the planted approvals 6 a run (the base a third of the hand's where the state holds three audit seats in four), F̂^A 0.59–0.76 read and never acted on (ten adjudications take two runs), m^A 1.00; the member-confirmed channel carries the excess still; §8.6: the audit layer is the state's and is paid as the state's (its review fees 2.2× the hand's). The liar chain: the false approvals 2.7× / 2.4× the hand's from 5.2× / 5.0×, the lying appeal paid 5–6 in a hundred from 23–31, the planted lie no longer told apart (0.67, composition's level), the lie dial reading the seat (m^La 2.7–3.9); (B) alone carries the whole of it, (A) alone reads the careful basin.

Source: HANDOFF.md l.2478–2478

a learning half beside a hand half settled at care 0.70

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

Run 41 (notes §71): the half on the corrected engine — where the members hold the audit seats the lever moves (the base at the hand's, the planted approvals 2.5× the hand's by the run's end, m^A 4.5–7.6, the direct's audit stamp −0.19), the standing false approvals 1.5–1.8× the hand's from 1.6–2.0× (−4.0 ± 1.7 / −2.4 ± 1.6 paired), the member-confirmed audit the whole excess, the state's fee income at the hand's

Source: HANDOFF.md l.2478–2478

the learned populations' attributed reward was a pooled purse — every slot's events credited to the latest decision on the case whatever slot's wallet they touched, a learner worker's price vesting onto the learner auditor's decision; runs 34–41 trained on it — corrected as own-name credit behind a flag

Source: HANDOFF.md l.2478–2478

2.8.0 and 2.9.0 stand as rules.

Source: HANDOFF.md l.2478–2478

Run 42 — the careful basin at 2.9.0 retrained on own-name credit (learned-rev-careful-290-own/): the careful population's on every measure but the standing false approvals — care 0.84, the initial seats stamped 1–5%, the audit seat 0.45–0.66 (the recompute's half, blind to the case), the bad plants approved 19 / 12, the suspensions 5, the members holding their audit seats and the lever moving (F̂^A 0.45 / 0.61, m^A 2.7), wealth, price, tasks per credit, the cost of truth and the fee split at the careful's and the hand's; the standing false approvals 28.7 / 18.5 (2.8× / 1.8× the hand's) carried by the never-audited channel at odds 0.28 (17.6 / 8.7 against the hand's 3.5) and the audit seat's half-stamp (10.2 / 9.1 against 2.9)

Source: HANDOFF.md l.2478–2478

The basin test read (November 6, 2026; notes §73): the 2.9.0 stamper start lands in run 42's basin

Source: HANDOFF.md l.2478–2478

one basin from both starts, both learners

Source: HANDOFF.md l.2478–2478

The liar chain on own-name credit read (notes §73): the careful's — care 0.80, the panels stamped 0.04 / 0.004 (from 0.21–0.26), the lying appeal paid 2.2% / 1.6% against the hand's 2.7% (the shared-purse 5–6%, 2.8.0's 23–31%)

Source: HANDOFF.md l.2478–2478

The roll's defaults move (the same ruling, verbatim): the halving window to 20 and p_floor to 0.5, as Appendix A defaults under §19 (not a rule change), since together they close the never-audited channel 17.6 → 1.8 at the price of 18% more audits on the members' seats — that cost read on the honest populations in §8.3 before consolidation, the window retuned first if it runs too high — and the seat's remaining half-stamp is left to 2.9.0's lever, which is still moving.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1541–1541

the never-audited standing false approvals 17.6 → 1.8 a run on the direct, 8.7 → 1.3 on the head alone, the hand's own 3.5 → 0, the odds at the verdict of what is left 0.28 → 0.55, the window the stronger knob; the audit seat's half-stamp untouched by the odds (member-confirmed +7.3 to +9.0 / +5.2 to +6.3 over the hand in every cell); the audits a run +18% on the learned population and +16% on the hand's, the members' seats carrying it.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1541–1541

**§8.3 / §8.6 at the new defaults read (November 6, 2026; notes §75; 2.9.0-s83/): the defaults' cost on the honest populations is a quarter to a third more audit seats on the members (+297 / +83 / +318 / +319 a run in §56's four cells; the state's seats unmoved; audits of good work), the care on real cases +7–10%, the cost of truth +3–5%, paid at 1–2% of the customers' tasks per credit with the members and Ω unmoved, and the honest populations' own standing false approvals down a quarter (60/40) to two thirds (all-careful)

Source: HANDOFF.md l.2478–2478

a seed's episode depends on the platform through the emission's pow alone

Source: HANDOFF.md l.2478–2478

every run from it reproducing anywhere

Source: HANDOFF.md l.2478–2478

Read (notes §77): the seam alone is noise (the 2.7.0 branch's pairing within error; the hand at the old defaults +0.6 ± 0.9 false approvals); at the consolidated defaults run 42 reads 16.3 / 9.9 standing false approvals against the hand's 7.2 (2.3× / 1.4×), the never-audited channel closed, the audit seat's half-stamp the excess; the 2.7.0 pairing on one episode keeps §73's sentence (the audit seat 0.94–0.97 → 0.41–0.56)

Source: HANDOFF.md l.2478–2478

A standing false approval: a real root (not a plant, not the state's delivery) vested with its work unfulfilled.

Source: agentic_sim/results/redteam/stability/cr-finals/fa-channels-careful-290-own-cr.md l.1–1

policy	standing false approvals	member confirmed	state confirmed	audited, rejected, vested	disputed, panel confirmed (a member sat)	disputed, panel confirmed (the state's alone)	appealed, panel sided with the appellant (a member sat)	appealed, panel sided with the appellant (the state's alone)	never audited	never-audited share	odds at verdict of the never-audited (mean)
direct	16.3 ± 1.2	13.2 ± 1.0	0.6 ± 0.2	0.5 ± 0.2	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	2.0 ± 0.4	0.123	0.518
head alone	9.9 ± 0.6	8.4 ± 0.6	0.5 ± 0.1	0.1 ± 0.1	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.9 ± 0.2	0.090	0.555
hand	6.9 ± 0.8	3.3 ± 0.4	2.9 ± 0.4	0.8 ± 0.2	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.000	nan

Source: agentic_sim/results/redteam/stability/cr-finals/fa-channels-careful-290-own-cr.md l.3–7

Conclusion

Not settled: the audit seat's remaining half-stamp is left to a lever that is still moving.

Seats reached only through dishonesty are now priced

Not settled: evaluations on the corrected engine and the half on own money remain

2.9.0 is the current rulebook, and its results are not settled. The record says the audit seat's remaining half is left to 2.9.0's lever, which is still moving. Several of the planned tests have not yet been run on the corrected setup. What the record does say: under 2.7.0 nothing put a price on a careless approval at the seats reached only through someone's dishonesty, and now something does. On the inspectors' own money, approving without looking at the audit seat fell from about nine times in ten to about half.

The record's rationale, restated by the reviewer: the seats a case reaches only through someone's dishonesty "were priced by nothing at 2.7.0 and are priced now, which is what they were for."

2.9.0's results are not settled. The record leaves the audit seat's remaining half-stamp to 2.9.0's lever, "which is still moving", and its next steps are still open: the remaining evaluations on the corrected engine and signal, then the half on own money.

The rationale, restated once (the reviewer's fourth ruling, verbatim): 2.8.0 and 2.9.0 stand on the own-money population as they stood on the pooled one — at 2.7.0 a self-play population of reviewers trained on their own money cares at the first seat, where a plant may come, and stamps the seats past it nine in ten, where none can, the audit layer captured and the false approvals three to four times an honest population's; the two revisions put the orchestrator's known truth at those seats, and from both starts the audit seat's stamp falls from 0.92 to 0.56 with the lever reading it — the seats a case reaches only through someone's dishonesty were priced by nothing at 2.7.0 and are priced now, which is what they were for.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1541–1541

The window's retune the reviewer's (the implementer would leave it). Next: §8.4–§8.7 on the corrected engine and signal, then the half on own money.

Source: HANDOFF.md l.2478–2478